Abstention on SPACEREJECT 135 ID + 135 OOD (held-out seq 3, 4, 5)
95.63Balanced AccuracySEMANTIC FLIP, Q-FLIP+V-FLIP
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SEMANTIC FLIP, Q-FLIP+V-FLIPBackbone=Qwen2.5-VL-7B, Protocol=learned MLP gate2026.06 | 95.63 | 95.59 | 94.67 | 96.59 | |
| SEMANTIC FLIP, Q-FLIP onlyBackbone=Qwen2.5-VL-7B, Protocol=learned MLP gate2026.06 | 95.04 | 94.94 | 93.63 | 96.44 | |
| C2 (Tool)Backbone=Qwen3-8B, Protocol=prompt-based2026.06 | 87.78 | 88.74 | 96.3 | 79.26 | |
| SEMANTIC FLIP, V-FLIP onlyBackbone=Qwen2.5-VL-7B, Protocol=learned MLP gate2026.06 | 72.37 | 68.67 | 63.85 | 80.89 |